[Mesa-dev] [PATCH] i965: Query the actual GTT size for reporting total usuable GPU memory

[Daniel Vetter]

On Mon, Oct 19, 2015 at 12:08:41PM +0100, Chris Wilson wrote:
> The amount of memory that the GPU can use in any single batch is the
> amount of memory it can map into its page tables, that is the GTT size.
> With per-process GTT (i.e. every context has its own set of page tables),
> this value may differ from the global GTT size (and is often larger).
> When reporting how much memory a client may use with the GPU, it is the
> size of the context GTT that is the limiting factor.
> 
> (Additionally, get_aperture_ioctl is relatively expensive for our query
> of the constant GTT size as it also computes addition information not
> used here and has the side-effect of doing a sysfs scan for PCI
> devices.)
> 
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>

Since I just pulled in the kernel side earlier today:

Reviewed-by: Daniel Vetter <daniel.vetter at ffwll.ch>

> Cc: Daniel Vetter <daniel at ffwll.ch>
> Cc: Kristian Høgsberg <krh at bitplanet.net>
> Cc: Kenneth Graunke <kenneth at whitecape.org>
> Cc: Ian Romanick <ian.d.romanick at intel.com>
> ---
>  src/mesa/drivers/dri/i965/intel_screen.c | 22 +++++++++++++++++-----
>  1 file changed, 17 insertions(+), 5 deletions(-)
> 
> diff --git a/src/mesa/drivers/dri/i965/intel_screen.c b/src/mesa/drivers/dri/i965/intel_screen.c
> index b77b747..cdf5b0d 100644
> --- a/src/mesa/drivers/dri/i965/intel_screen.c
> +++ b/src/mesa/drivers/dri/i965/intel_screen.c
> @@ -787,6 +787,22 @@ static const __DRIimageExtension intelImageExtension = {
>      .getCapabilities                    = NULL
>  };
>  
> +static uint64_t intel_get_gtt_size(int fd)
> +{
> +   struct drm_i915_gem_context_param p;
> +   size_t mappable_size, aper_size;
> +
> +   memset(&p, 0, sizeof(p));
> +#define I915_CONTEXT_PARAM_GTT_SIZE 0x3 /* XXX requires future libdrm */
> +   p.param = I915_CONTEXT_PARAM_GTT_SIZE;

Optional bikeshed: Maybe explicitly set the context to 0 to make it clear
we're querying the (always present) default context here.

> +   if (drmIoctl(fd, DRM_IOCTL_I915_GEM_CONTEXT_GETPARAM, &p) == 0)
> +      return p.value;
> +
> +   drm_intel_get_aperture_sizes(fd, &mappable_size, &aper_size);
> +
> +   return aper_size;
> +}
> +
>  static int
>  brw_query_renderer_integer(__DRIscreen *psp, int param, unsigned int *value)
>  {
> @@ -808,11 +824,7 @@ brw_query_renderer_integer(__DRIscreen *psp, int param, unsigned int *value)
>         * assume that there's some fragmentation, and we start doing extra
>         * flushing, etc.  That's the big cliff apps will care about.
>         */
> -      size_t aper_size;
> -      size_t mappable_size;
> -
> -      drm_intel_get_aperture_sizes(psp->fd, &mappable_size, &aper_size);
> -
> +      uint64_t aper_size = intel_get_gtt_size(psp->fd);
>        const unsigned gpu_mappable_megabytes =
>           (aper_size / (1024 * 1024)) * 3 / 4;
>  
> -- 
> 2.6.1
> 

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch